Run end-to-end latency benchmark on burst workload

[AayushBarhate]

What

Run benchmark/latency/run_latency_test.sh against real Kafka burst traffic. Collect per-event latency for fixed (2s), fixed (500ms), and adaptive approaches.

Why

This is the paper's main result. Benchmark code exists but hasn't been executed with burst data flowing through Kafka → Spark.

Steps

1. Start Kafka, create topic adaptive-latency-test
2. Run generator/burst_generator.py — 300s, baseline=100, burst_mult=8
3. Run benchmark/latency/latency_benchmark.py for each mode
4. Collect JSON from /tmp/latency_results/
5. Compute p50, p95, p99 per approach
6. Adaptive should beat fixed during burst windows

Acceptance criteria

• Latency JSON for all 3 approaches on same workload
• Comparison table showing adaptive reduces burst-period latency